When Innovation Becomes Narrative
Artificial intelligence promised to change drug discovery. The bottleneck moved.
In October 2023 a company announced it was discontinuing its lead clinical program.
The molecule had been celebrated as proof that artificial intelligence had fundamentally changed drug discovery.
It was a Phase 1 failure.
The company had been valued at $2.9 billion eighteen months earlier.
The molecule wasn’t the problem.
The problem was what the molecule was supposed to prove — and didn’t.
Artificial intelligence has entered drug discovery with extraordinary fanfare and extraordinary capital.
It has not yet produced an extraordinary drug.
That gap is where this note lives.
The bottleneck moves
Drug discovery has a long history of tools that changed everything.
And a long history of attrition curves that didn’t move.
The Human Genome Project delivered the complete human DNA sequence in 2003.
It hasn’t delivered a proportional increase in approved drugs.
High throughput screening promised to identify drug candidates faster than prior methods.
It delivered improved hit rates but clinical success rates didn’t follow.
Combinatorial chemistry promised to expand the universe of testable molecules exponentially.
It delivered compounds too synthetic to translate biologically and the approach was largely abandoned.
RNAi promised to silence any gene in the human genome with unprecedented precision.
It delivered a modality that took fifteen years to produce approved drugs.
The first wave of pure-play RNAi companies was largely destroyed in the interim.
Alnylam survived. Most didn’t.
Each tool promised to change drug discovery.
Each tool removed a genuine bottleneck.
Each time the bottleneck moved.
The pipeline remained as long as it was before.
The fully loaded cost of drug development has continued to rise through every one of these waves.
Not because the science failed.
But because solving one constraint reveals the next one.
Artificial intelligence is the latest tool to enter this sequence.
It has genuine capabilities prior tools did not.
It has also encountered the same structural reality every prior tool encountered.
The bottleneck moved.
It is now located precisely where the artificial intelligence platforms are not.
The kiss of death is not failure.
It is success that arrives from somewhere else.
Millennium Pharmaceuticals built the defining genomics platform of its era.
Its one commercially significant drug came from an acquisition.
Velcade had nothing to do with the platform.
The platform didn’t deliver the pipeline.
The acquisition did.
The market noticed. Eventually.
What would it take?
Before examining the companies, the evidentiary bar deserves to be stated explicitly.
What would it actually take for artificial intelligence drug discovery to justify current valuations?
Not process metrics.
Not partnership announcements.
Not platform demonstrations.
We’ll propose one here.
An artificial intelligence-native molecule — discovered, optimized, and advanced because of the platform, not despite it — reaching Phase 2 with clean data.
Specifically —
Partnership economics that reflect drug value rather than platform access.
Milestones tied to clinical outcomes, not signatures.
A cost-per-approved-drug that demonstrably beats the industry average.
Not one molecule. A pattern.
None of those bars have been cleared.
That is not an opinion.
It is a current scorecard.
The market has not established what the evidentiary standard should be.
That is not an accident because —
Unfalsifiable promises are more fundable than falsifiable ones.
A process milestone — molecules identified faster, hit rates improved, screening costs reduced — can always be presented as progress.
It is not the same as proof.
The distinction between process and outcome is where most of the valuation work in artificial intelligence drug discovery is currently being done.
Quietly.
Without acknowledgment.
The HGP era had the same architecture.
Sequence coverage was the process metric.
Approved drugs were the outcome metric.
The market priced sequence coverage as if it were approved drugs.
It wasn’t.
Artificial intelligence drug discovery is pricing platform capability as if it were clinical validation.
It isn’t.
Not yet.
Perhaps not for a long time.
Recursion Pharmaceuticals
Recursion Pharmaceuticals is the largest pure-play artificial intelligence drug discovery company in public markets.
It is also the most instructive.
The company maps biological relationships at scale using automated microscopy and machine learning — millions of experiments generating billions of data points about how cells respond to genetic and chemical perturbations.
The platform is real.
The data are real.
The $50 million investment Nvidia made in 2023 is real.
The market read it as scientific validation.
But, to us, it looks a lot like a commercial relationship between a chipmaker and one of its largest compute customers.
Those are different things.
The pipeline tells the more important story.
REC-994 for symptomatic cerebral cavernous malformation — discontinued.
REC-2282 for progressive neurofibromatosis type II — discontinued.
REC-3964 for prevention of C. difficile infection — discontinued.
The surviving clinical programs include assets absorbed through the acquisition of Exscientia.
Several carry new names.
The compounds are the same.
Success arriving from somewhere else.
Current market cap approximately $2 billion.
Net of ~$665 million cash, enterprise value assigned to a surviving pipeline where the lead program faces regulatory discussions not expected until the second half of 2026 and additional Phase 1b/2 data not until the first half of 2027.
The platform generates enormous amounts of biological data.
The question the valuation requires you to answer is whether data generation is drug discovery.
Or whether it is a necessary but nowhere near sufficient condition for it.
Millennium generated enormous amounts of genomic data.
Velcade came from LeukoSite.
“I believe the combination of the incredible teams and platforms at Exscientia and Recursion position us as the leader of the AI-enabled drug discovery and development space.”
— Chris Gibson, CEO, Recursion Pharmaceuticals
Our mission is to Decode Biology to Radically Improve Lives.
— Recursion Pharmaceuticals, Dear Investor
The flywheel has more than $20 billion in possible milestone payments.
It has received $450 million.
BioBucks.
By contrast, consider Genzyme circa 1991.
One asset.
Ceredase for Gaucher disease.
One enzyme replacement therapy for one rare disease with a defined patient population and a clear biological rationale.
No flywheel.
No BioBucks.
No mission to decode biology.
A drug that worked.
That single asset built a multi-billion-dollar enterprise.
Not because of what Genzyme promised.
Because of what Genzyme delivered.
The biotech sector pipeline chart has changed considerably in thirty years.
The standard for what constitutes proof shouldn’t.
Absci
Absci is the hardest of the three to pin down.
That is itself an analytical observation.
The company applies generative artificial intelligence to antibody design.
The pitch sits at the intersection of two hype cycles simultaneously.
Artificial intelligence. Biologics drug discovery.
That is a lot of narrative capital trying to do a lot of valuation work.
The numbers are stark.
Current market cap approximately $1 billion.
Trailing twelve-month revenue — $1.84 million.
A revenue multiple of roughly 540 times.
For context, a revenue multiple requires revenue — the $1.84 million is not revenue in any conventional commercial sense.
It is the accounting classification applied to partnership payments received before any drug has been discovered, developed, or delivered.
The clinical pipeline consists of two Phase 1 assets and two preclinical programs.
Interim proof of concept data for the lead program is not expected until the second half of 2026.
A Phase 2 trial for the second program does not start until late 2026 with interim data not until the second half of 2027.
Investors are being asked to look out eighteen to twenty-four months, minimum, for any outcome signal.
At a valuation that prices in significant success.
The what would it take question is particularly stark for Absci.
The answer is years away by any honest development timeline.
The current valuation requires you to believe the answer is already known.
It is not.
The generative artificial intelligence framing deserves specific attention.
Generative artificial intelligence produces novel antibody sequences that satisfy specified design criteria.
It does not produce drugs.
The distance between a computationally designed antibody sequence and an approved therapeutic is the entire drug development enterprise.
Clinical trials. Regulatory review. Manufacturing. Commercial execution.
The artificial intelligence handles the first step.
The bottleneck moved.
The generative artificial intelligence capability may be real.
The clinical development infrastructure required to convert it into an approved drug is a separate question.
One the current management team has not yet been required to answer publicly.
Insilico Medicine
Insilico Medicine is the sector’s own best argument for itself.
Which makes it the most interesting case to put through the framework.
The company went public on the Hong Kong Stock Exchange in December 2025.
It raised approximately $293 million at an implied valuation of $1.6 to $1.7 billion.
The pipeline is the most mature of the three companies.
Seven programs in active clinical development.
Thirty plus programs in total.
Full year 2025 revenue of $56 million — the highest of the three by a significant margin.
The bull case is straightforward.
More programs. More revenue. More clinical activity. More proof.
The bear case requires only one question.
What is the revenue actually measuring?
The $56 million is substantially derived from software licensing and discovery services.
Not drug value.
A biotech company that licenses its discovery platform to others is a software company with a biology wrapper.
The valuation it deserves is a software valuation.
Not a therapeutic company valuation.
The distinction matters enormously.
A therapeutic company valuation prices the option on approved drugs and commercial revenues.
A software valuation prices recurring license fees and service contracts.
Those are different multiples applied to different cash flow streams.
Insilico is currently valued as the former while generating revenue that looks more like the latter.
The pipeline is real.
The clinical activity is real.
The revenue is real.
The question is whether any of it clears the bar the sector set for itself.
An artificial intelligence-native molecule discovered, optimized, and advanced because of the platform.
Reaching Phase 2 with clean data.
A pattern not a single data point.
Insilico has the most extensive pipeline of the three.
It has not yet produced that molecule.
The sector’s best argument remains an argument.
Not yet a proof.
What the disclosures show
The three companies examined here are public.
Their disclosures are available to anyone willing to read them carefully.
Most don’t.
Partnership economics deserve specific attention.
A $500 million collaboration announcement is not a $500 million payment.
It is BioBucks.
(A term of art that deserves wider circulation in the sector.)
A modest upfront fee attached to a long tail of milestones contingent on outcomes that may never occur.
The headline numbers do the narrative work.
The footnotes contain the actual economics.
Sanofi’s $5.2 billion Exscientia partnership is the canonical example.
Fifteen programs. All milestones. Maximum theoretical value.
The molecule that actually entered clinical trials failed in Phase 1.
The BioBucks never became dollars.
Cash runway versus data timeline is simple arithmetic.
How long does current cash last at current burn rate?
When does the next meaningful clinical readout occur?
If the runway ends before the proof point arrives the company faces a dilutive raise at exactly the moment the narrative is most vulnerable.
That math is in every quarterly filing.
Almost nobody does it.
Endpoint selection tells an experienced analyst whether a company expects to generate clean data or is designing for ambiguity.
A primary endpoint that is a biomarker surrogate rather than a clinical outcome is a tell.
It is not evidence of fraud — it is evidence of a company managing the distance between what the platform can demonstrate and what the market requires it to demonstrate.
An observation that deserves repeating.
It is not evidence of fraud — it shines a bright light on a company managing the gap between what the platform can deliver and what the market is pricing in occurring.
Partnership attrition is the most underappreciated signal.
Announced partnerships generate press releases.
Discontinued partnerships generate quiet 8-K filings that nobody reads.
The ratio of announced to discontinued tells you more about platform conviction than any investor presentation.
Exscientia’s Bayer partnership — €240 million announced, quietly terminated — was in the public record the entire time.
The choreography just didn’t feature it prominently.
The financing cycle
Artificial intelligence drug discovery companies share a common financial architecture.
No drug revenues.
Significant cash burn.
Survival dependent on continuous capital formation.
That is not a criticism.
It is a description of early stage biotech broadly.
The difference is what justifies the next raise.
In conventional drug development the next raise is justified by clinical progress.
A Phase 1 readout. A Phase 2 enrollment milestone. A regulatory designation.
Evidence that the program is advancing toward the outcome that justifies the valuation.
In artificial intelligence drug discovery the next raise is increasingly justified by platform metrics.
Compounds screened. Models trained. Partnerships announced.
Process milestones presented as evidence of progress toward outcomes they do not yet predict.
The financing window requires a story.
The story requires a catalyst.
The catalyst is whatever the platform can currently demonstrate.
Not whatever the market should require it to demonstrate.
The BioBucks partnership is the most reliable financing catalyst in the sector.
A large pharma name attached to a large headline number resets the narrative clock.
It signals external validation without requiring external proof.
It buys eighteen to twenty-four months of runway before the next question arrives.
By which point another partnership can be announced.
Or another acquisition can reset the clock entirely.
Recursion absorbing Exscientia is the cleanest example.
Two companies unable to demonstrate a connection between platform capability and clinical output.
Combined into one larger platform.
With a new narrative about synergy.
And a new runway before the same question arrives again.
This is not Minsky drift.
The cycle wasn’t an accident that emerged from excess.
It was the architecture from the beginning.
The gap between capability and proof is not a problem to be solved.
It is the substrate on which the financing cycle runs.
Managed carefully it can run for a very long time.
The tada moment would actually end it.
The distinction is worth stating precisely.
Screening compounds and sequencing genes has value.
It is the value a contract research organization provides.
CROs are priced accordingly.
What artificial intelligence drug discovery platforms have underwritten is something different.
Durable repeatable value creation.
A platform that reliably generates clinical candidates.
That advance through development.
That produce approved drugs.
At a rate and cost that justifies a therapeutic company valuation.
That is the implicit promise in every investor letter.
It is not what the scoreboard currently shows.
Process metrics are not platform validation.
They are evidence that the platform is running.
Not evidence that it is working.
Getting it right
The fair question is whether any company in the sector is getting it right.
Or at least not a trainwreck candidate.
The honest answer is that the evidentiary bar hasn’t been cleared by anyone yet.
The tada moment remains elusive across the board.
Some companies have approached the problem more circumspectly.
Schrodinger applies physics-based computational modeling to predict molecular binding and behavior.
The value proposition is specific and falsifiable.
Not, we use artificial intelligence to discover drugs.
Rather — our models predict which compounds are worth synthesizing before you synthesize them.
That is a more honest claim about what computation can actually do in drug discovery.
It is also a lower claim than most of the sector makes about itself.
Schroeder has yet to produce the clean Phase 2 readout that would constitute proof of concept for the platform.
But the methodology is more honest about where the bottleneck actually sits.
Computational approaches that make specific falsifiable claims about what they can do deserve more analytical credit than platforms that promise to decode biology.
And more skepticism than an approved drug would warrant.
The sector exists on a spectrum.
Most of the capital has not concentrated at the honest end of it.
The loop closes
The company in the cold open was Exscientia.
The molecule was EXS21546.
Developed for renal cell carcinoma and non-small cell lung cancer.
Celebrated as the first artificial intelligence-designed molecule to reach clinical trials.
Discontinued in Phase 1 in October 2023.
The company had been valued at $2.9 billion at its IPO in October 2021.
It was acquired by Recursion in 2024 for approximately $630 million.
A 78% decline from peak valuation.
The Gates Foundation invested at IPO.
Sanofi signed a partnership carrying up to $5.2 billion in potential milestones.
For up to fifteen novel small molecule candidates across oncology and immunology.
Potential milestones if all milestones for all programs are achieved is not a valuation — it is an option portfolio priced at maximum theoretical value.
The market treated it as a valuation.
That gap — between option theory and expected value — is where narrative capital lives.
The molecule wasn’t the problem.
The problem was what the molecule was supposed to prove.
And didn’t.
The platform generated the molecule.
The molecule failed in Phase 1.
The company was absorbed into another company whose own platform has discontinued three programs and sourced its most advanced remaining assets from the acquisition.
Success arrived from somewhere else.
Twice.
The choreography continues.
The clock resets.
The narrative capital raises again.
Under a different name.
With a larger platform.
And the same distance between capability and proof.
Coda
Artificial intelligence has completed the journey from technology to buzzword.
The underlying capability remains real.
The language no longer requires examination to deploy.
That is when the mispricing begins.
Every cycle has its vocabulary.
Genomics. Proteomics. Combinatorial chemistry. RNAi. Synthetic biology.
Each one arrived with genuine capability and genuine promise.
Each one became ambient language before the proof arrived.
The buzzword doesn’t kill the technology.
It kills the analytical standard applied to the technology.
When artificial intelligence appears in every investor letter, every partnership announcement, every pipeline description, every mission statement —
It has stopped being a description.
It has become a signal.
Sophisticated. Current. Unfalsifiable.
The market prices the signal.
The patients wait for the drug.
The distance between those two things is where Mispriced Biology lives.